Loading...
Note: File does not exist in v3.1.
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * (C) 2001 Clemson University and The University of Chicago
4 * Copyright 2018 Omnibond Systems, L.L.C.
5 *
6 * See COPYING in top-level directory.
7 */
8
9/*
10 * Linux VFS file operations.
11 */
12
13#include "protocol.h"
14#include "orangefs-kernel.h"
15#include "orangefs-bufmap.h"
16#include <linux/fs.h>
17#include <linux/pagemap.h>
18
19static int flush_racache(struct inode *inode)
20{
21 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
22 struct orangefs_kernel_op_s *new_op;
23 int ret;
24
25 gossip_debug(GOSSIP_UTILS_DEBUG,
26 "%s: %pU: Handle is %pU | fs_id %d\n", __func__,
27 get_khandle_from_ino(inode), &orangefs_inode->refn.khandle,
28 orangefs_inode->refn.fs_id);
29
30 new_op = op_alloc(ORANGEFS_VFS_OP_RA_FLUSH);
31 if (!new_op)
32 return -ENOMEM;
33 new_op->upcall.req.ra_cache_flush.refn = orangefs_inode->refn;
34
35 ret = service_operation(new_op, "orangefs_flush_racache",
36 get_interruptible_flag(inode));
37
38 gossip_debug(GOSSIP_UTILS_DEBUG, "%s: got return value of %d\n",
39 __func__, ret);
40
41 op_release(new_op);
42 return ret;
43}
44
45/*
46 * Post and wait for the I/O upcall to finish
47 */
48ssize_t wait_for_direct_io(enum ORANGEFS_io_type type, struct inode *inode,
49 loff_t *offset, struct iov_iter *iter, size_t total_size,
50 loff_t readahead_size, struct orangefs_write_range *wr,
51 int *index_return, struct file *file)
52{
53 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
54 struct orangefs_khandle *handle = &orangefs_inode->refn.khandle;
55 struct orangefs_kernel_op_s *new_op = NULL;
56 int buffer_index;
57 ssize_t ret;
58 size_t copy_amount;
59 int open_for_read;
60 int open_for_write;
61
62 new_op = op_alloc(ORANGEFS_VFS_OP_FILE_IO);
63 if (!new_op)
64 return -ENOMEM;
65
66 /* synchronous I/O */
67 new_op->upcall.req.io.readahead_size = readahead_size;
68 new_op->upcall.req.io.io_type = type;
69 new_op->upcall.req.io.refn = orangefs_inode->refn;
70
71populate_shared_memory:
72 /* get a shared buffer index */
73 buffer_index = orangefs_bufmap_get();
74 if (buffer_index < 0) {
75 ret = buffer_index;
76 gossip_debug(GOSSIP_FILE_DEBUG,
77 "%s: orangefs_bufmap_get failure (%zd)\n",
78 __func__, ret);
79 goto out;
80 }
81 gossip_debug(GOSSIP_FILE_DEBUG,
82 "%s(%pU): GET op %p -> buffer_index %d\n",
83 __func__,
84 handle,
85 new_op,
86 buffer_index);
87
88 new_op->uses_shared_memory = 1;
89 new_op->upcall.req.io.buf_index = buffer_index;
90 new_op->upcall.req.io.count = total_size;
91 new_op->upcall.req.io.offset = *offset;
92 if (type == ORANGEFS_IO_WRITE && wr) {
93 new_op->upcall.uid = from_kuid(&init_user_ns, wr->uid);
94 new_op->upcall.gid = from_kgid(&init_user_ns, wr->gid);
95 }
96 /*
97 * Orangefs has no open, and orangefs checks file permissions
98 * on each file access. Posix requires that file permissions
99 * be checked on open and nowhere else. Orangefs-through-the-kernel
100 * needs to seem posix compliant.
101 *
102 * The VFS opens files, even if the filesystem provides no
103 * method. We can see if a file was successfully opened for
104 * read and or for write by looking at file->f_mode.
105 *
106 * When writes are flowing from the page cache, file is no
107 * longer available. We can trust the VFS to have checked
108 * file->f_mode before writing to the page cache.
109 *
110 * The mode of a file might change between when it is opened
111 * and IO commences, or it might be created with an arbitrary mode.
112 *
113 * We'll make sure we don't hit EACCES during the IO stage by
114 * using UID 0. Some of the time we have access without changing
115 * to UID 0 - how to check?
116 */
117 if (file) {
118 open_for_write = file->f_mode & FMODE_WRITE;
119 open_for_read = file->f_mode & FMODE_READ;
120 } else {
121 open_for_write = 1;
122 open_for_read = 0; /* not relevant? */
123 }
124 if ((type == ORANGEFS_IO_WRITE) && open_for_write)
125 new_op->upcall.uid = 0;
126 if ((type == ORANGEFS_IO_READ) && open_for_read)
127 new_op->upcall.uid = 0;
128
129 gossip_debug(GOSSIP_FILE_DEBUG,
130 "%s(%pU): offset: %llu total_size: %zd\n",
131 __func__,
132 handle,
133 llu(*offset),
134 total_size);
135 /*
136 * Stage 1: copy the buffers into client-core's address space
137 */
138 if (type == ORANGEFS_IO_WRITE && total_size) {
139 ret = orangefs_bufmap_copy_from_iovec(iter, buffer_index,
140 total_size);
141 if (ret < 0) {
142 gossip_err("%s: Failed to copy-in buffers. Please make sure that the pvfs2-client is running. %ld\n",
143 __func__, (long)ret);
144 goto out;
145 }
146 }
147
148 gossip_debug(GOSSIP_FILE_DEBUG,
149 "%s(%pU): Calling post_io_request with tag (%llu)\n",
150 __func__,
151 handle,
152 llu(new_op->tag));
153
154 /* Stage 2: Service the I/O operation */
155 ret = service_operation(new_op,
156 type == ORANGEFS_IO_WRITE ?
157 "file_write" :
158 "file_read",
159 get_interruptible_flag(inode));
160
161 /*
162 * If service_operation() returns -EAGAIN #and# the operation was
163 * purged from orangefs_request_list or htable_ops_in_progress, then
164 * we know that the client was restarted, causing the shared memory
165 * area to be wiped clean. To restart a write operation in this
166 * case, we must re-copy the data from the user's iovec to a NEW
167 * shared memory location. To restart a read operation, we must get
168 * a new shared memory location.
169 */
170 if (ret == -EAGAIN && op_state_purged(new_op)) {
171 orangefs_bufmap_put(buffer_index);
172 if (type == ORANGEFS_IO_WRITE)
173 iov_iter_revert(iter, total_size);
174 gossip_debug(GOSSIP_FILE_DEBUG,
175 "%s:going to repopulate_shared_memory.\n",
176 __func__);
177 goto populate_shared_memory;
178 }
179
180 if (ret < 0) {
181 if (ret == -EINTR) {
182 /*
183 * We can't return EINTR if any data was written,
184 * it's not POSIX. It is minimally acceptable
185 * to give a partial write, the way NFS does.
186 *
187 * It would be optimal to return all or nothing,
188 * but if a userspace write is bigger than
189 * an IO buffer, and the interrupt occurs
190 * between buffer writes, that would not be
191 * possible.
192 */
193 switch (new_op->op_state - OP_VFS_STATE_GIVEN_UP) {
194 /*
195 * If the op was waiting when the interrupt
196 * occurred, then the client-core did not
197 * trigger the write.
198 */
199 case OP_VFS_STATE_WAITING:
200 if (*offset == 0)
201 ret = -EINTR;
202 else
203 ret = 0;
204 break;
205 /*
206 * If the op was in progress when the interrupt
207 * occurred, then the client-core was able to
208 * trigger the write.
209 */
210 case OP_VFS_STATE_INPROGR:
211 if (type == ORANGEFS_IO_READ)
212 ret = -EINTR;
213 else
214 ret = total_size;
215 break;
216 default:
217 gossip_err("%s: unexpected op state :%d:.\n",
218 __func__,
219 new_op->op_state);
220 ret = 0;
221 break;
222 }
223 gossip_debug(GOSSIP_FILE_DEBUG,
224 "%s: got EINTR, state:%d: %p\n",
225 __func__,
226 new_op->op_state,
227 new_op);
228 } else {
229 gossip_err("%s: error in %s handle %pU, returning %zd\n",
230 __func__,
231 type == ORANGEFS_IO_READ ?
232 "read from" : "write to",
233 handle, ret);
234 }
235 if (orangefs_cancel_op_in_progress(new_op))
236 return ret;
237
238 goto out;
239 }
240
241 /*
242 * Stage 3: Post copy buffers from client-core's address space
243 */
244 if (type == ORANGEFS_IO_READ && new_op->downcall.resp.io.amt_complete) {
245 /*
246 * NOTE: the iovector can either contain addresses which
247 * can futher be kernel-space or user-space addresses.
248 * or it can pointers to struct page's
249 */
250
251 copy_amount = new_op->downcall.resp.io.amt_complete;
252
253 ret = orangefs_bufmap_copy_to_iovec(iter, buffer_index,
254 copy_amount);
255 if (ret < 0) {
256 gossip_err("%s: Failed to copy-out buffers. Please make sure that the pvfs2-client is running (%ld)\n",
257 __func__, (long)ret);
258 goto out;
259 }
260 }
261 gossip_debug(GOSSIP_FILE_DEBUG,
262 "%s(%pU): Amount %s, returned by the sys-io call:%d\n",
263 __func__,
264 handle,
265 type == ORANGEFS_IO_READ ? "read" : "written",
266 (int)new_op->downcall.resp.io.amt_complete);
267
268 ret = new_op->downcall.resp.io.amt_complete;
269
270out:
271 if (buffer_index >= 0) {
272 orangefs_bufmap_put(buffer_index);
273 gossip_debug(GOSSIP_FILE_DEBUG,
274 "%s(%pU): PUT buffer_index %d\n",
275 __func__, handle, buffer_index);
276 buffer_index = -1;
277 }
278 op_release(new_op);
279 return ret;
280}
281
282int orangefs_revalidate_mapping(struct inode *inode)
283{
284 struct orangefs_inode_s *orangefs_inode = ORANGEFS_I(inode);
285 struct address_space *mapping = inode->i_mapping;
286 unsigned long *bitlock = &orangefs_inode->bitlock;
287 int ret;
288
289 while (1) {
290 ret = wait_on_bit(bitlock, 1, TASK_KILLABLE);
291 if (ret)
292 return ret;
293 spin_lock(&inode->i_lock);
294 if (test_bit(1, bitlock)) {
295 spin_unlock(&inode->i_lock);
296 continue;
297 }
298 if (!time_before(jiffies, orangefs_inode->mapping_time))
299 break;
300 spin_unlock(&inode->i_lock);
301 return 0;
302 }
303
304 set_bit(1, bitlock);
305 smp_wmb();
306 spin_unlock(&inode->i_lock);
307
308 unmap_mapping_range(mapping, 0, 0, 0);
309 ret = filemap_write_and_wait(mapping);
310 if (!ret)
311 ret = invalidate_inode_pages2(mapping);
312
313 orangefs_inode->mapping_time = jiffies +
314 orangefs_cache_timeout_msecs*HZ/1000;
315
316 clear_bit(1, bitlock);
317 smp_mb__after_atomic();
318 wake_up_bit(bitlock, 1);
319
320 return ret;
321}
322
323static ssize_t orangefs_file_read_iter(struct kiocb *iocb,
324 struct iov_iter *iter)
325{
326 int ret;
327 orangefs_stats.reads++;
328
329 down_read(&file_inode(iocb->ki_filp)->i_rwsem);
330 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
331 if (ret)
332 goto out;
333
334 ret = generic_file_read_iter(iocb, iter);
335out:
336 up_read(&file_inode(iocb->ki_filp)->i_rwsem);
337 return ret;
338}
339
340static ssize_t orangefs_file_write_iter(struct kiocb *iocb,
341 struct iov_iter *iter)
342{
343 int ret;
344 orangefs_stats.writes++;
345
346 if (iocb->ki_pos > i_size_read(file_inode(iocb->ki_filp))) {
347 ret = orangefs_revalidate_mapping(file_inode(iocb->ki_filp));
348 if (ret)
349 return ret;
350 }
351
352 ret = generic_file_write_iter(iocb, iter);
353 return ret;
354}
355
356static vm_fault_t orangefs_fault(struct vm_fault *vmf)
357{
358 struct file *file = vmf->vma->vm_file;
359 int ret;
360 ret = orangefs_inode_getattr(file->f_mapping->host,
361 ORANGEFS_GETATTR_SIZE);
362 if (ret == -ESTALE)
363 ret = -EIO;
364 if (ret) {
365 gossip_err("%s: orangefs_inode_getattr failed, "
366 "ret:%d:.\n", __func__, ret);
367 return VM_FAULT_SIGBUS;
368 }
369 return filemap_fault(vmf);
370}
371
372static const struct vm_operations_struct orangefs_file_vm_ops = {
373 .fault = orangefs_fault,
374 .map_pages = filemap_map_pages,
375 .page_mkwrite = orangefs_page_mkwrite,
376};
377
378/*
379 * Memory map a region of a file.
380 */
381static int orangefs_file_mmap(struct file *file, struct vm_area_struct *vma)
382{
383 int ret;
384
385 ret = orangefs_revalidate_mapping(file_inode(file));
386 if (ret)
387 return ret;
388
389 gossip_debug(GOSSIP_FILE_DEBUG,
390 "orangefs_file_mmap: called on %pD\n", file);
391
392 /* set the sequential readahead hint */
393 vma->vm_flags |= VM_SEQ_READ;
394 vma->vm_flags &= ~VM_RAND_READ;
395
396 file_accessed(file);
397 vma->vm_ops = &orangefs_file_vm_ops;
398 return 0;
399}
400
401#define mapping_nrpages(idata) ((idata)->nrpages)
402
403/*
404 * Called to notify the module that there are no more references to
405 * this file (i.e. no processes have it open).
406 *
407 * \note Not called when each file is closed.
408 */
409static int orangefs_file_release(struct inode *inode, struct file *file)
410{
411 gossip_debug(GOSSIP_FILE_DEBUG,
412 "orangefs_file_release: called on %pD\n",
413 file);
414
415 /*
416 * remove all associated inode pages from the page cache and
417 * readahead cache (if any); this forces an expensive refresh of
418 * data for the next caller of mmap (or 'get_block' accesses)
419 */
420 if (file_inode(file) &&
421 file_inode(file)->i_mapping &&
422 mapping_nrpages(&file_inode(file)->i_data)) {
423 if (orangefs_features & ORANGEFS_FEATURE_READAHEAD) {
424 gossip_debug(GOSSIP_INODE_DEBUG,
425 "calling flush_racache on %pU\n",
426 get_khandle_from_ino(inode));
427 flush_racache(inode);
428 gossip_debug(GOSSIP_INODE_DEBUG,
429 "flush_racache finished\n");
430 }
431
432 }
433 return 0;
434}
435
436/*
437 * Push all data for a specific file onto permanent storage.
438 */
439static int orangefs_fsync(struct file *file,
440 loff_t start,
441 loff_t end,
442 int datasync)
443{
444 int ret;
445 struct orangefs_inode_s *orangefs_inode =
446 ORANGEFS_I(file_inode(file));
447 struct orangefs_kernel_op_s *new_op = NULL;
448
449 ret = filemap_write_and_wait_range(file_inode(file)->i_mapping,
450 start, end);
451 if (ret < 0)
452 return ret;
453
454 new_op = op_alloc(ORANGEFS_VFS_OP_FSYNC);
455 if (!new_op)
456 return -ENOMEM;
457 new_op->upcall.req.fsync.refn = orangefs_inode->refn;
458
459 ret = service_operation(new_op,
460 "orangefs_fsync",
461 get_interruptible_flag(file_inode(file)));
462
463 gossip_debug(GOSSIP_FILE_DEBUG,
464 "orangefs_fsync got return value of %d\n",
465 ret);
466
467 op_release(new_op);
468 return ret;
469}
470
471/*
472 * Change the file pointer position for an instance of an open file.
473 *
474 * \note If .llseek is overriden, we must acquire lock as described in
475 * Documentation/filesystems/locking.rst.
476 *
477 * Future upgrade could support SEEK_DATA and SEEK_HOLE but would
478 * require much changes to the FS
479 */
480static loff_t orangefs_file_llseek(struct file *file, loff_t offset, int origin)
481{
482 int ret = -EINVAL;
483 struct inode *inode = file_inode(file);
484
485 if (origin == SEEK_END) {
486 /*
487 * revalidate the inode's file size.
488 * NOTE: We are only interested in file size here,
489 * so we set mask accordingly.
490 */
491 ret = orangefs_inode_getattr(file->f_mapping->host,
492 ORANGEFS_GETATTR_SIZE);
493 if (ret == -ESTALE)
494 ret = -EIO;
495 if (ret) {
496 gossip_debug(GOSSIP_FILE_DEBUG,
497 "%s:%s:%d calling make bad inode\n",
498 __FILE__,
499 __func__,
500 __LINE__);
501 return ret;
502 }
503 }
504
505 gossip_debug(GOSSIP_FILE_DEBUG,
506 "orangefs_file_llseek: offset is %ld | origin is %d"
507 " | inode size is %lu\n",
508 (long)offset,
509 origin,
510 (unsigned long)i_size_read(inode));
511
512 return generic_file_llseek(file, offset, origin);
513}
514
515/*
516 * Support local locks (locks that only this kernel knows about)
517 * if Orangefs was mounted -o local_lock.
518 */
519static int orangefs_lock(struct file *filp, int cmd, struct file_lock *fl)
520{
521 int rc = -EINVAL;
522
523 if (ORANGEFS_SB(file_inode(filp)->i_sb)->flags & ORANGEFS_OPT_LOCAL_LOCK) {
524 if (cmd == F_GETLK) {
525 rc = 0;
526 posix_test_lock(filp, fl);
527 } else {
528 rc = posix_lock_file(filp, fl, NULL);
529 }
530 }
531
532 return rc;
533}
534
535static int orangefs_flush(struct file *file, fl_owner_t id)
536{
537 /*
538 * This is vfs_fsync_range(file, 0, LLONG_MAX, 0) without the
539 * service_operation in orangefs_fsync.
540 *
541 * Do not send fsync to OrangeFS server on a close. Do send fsync
542 * on an explicit fsync call. This duplicates historical OrangeFS
543 * behavior.
544 */
545 int r;
546
547 r = filemap_write_and_wait_range(file->f_mapping, 0, LLONG_MAX);
548 if (r > 0)
549 return 0;
550 else
551 return r;
552}
553
554/** ORANGEFS implementation of VFS file operations */
555const struct file_operations orangefs_file_operations = {
556 .llseek = orangefs_file_llseek,
557 .read_iter = orangefs_file_read_iter,
558 .write_iter = orangefs_file_write_iter,
559 .lock = orangefs_lock,
560 .mmap = orangefs_file_mmap,
561 .open = generic_file_open,
562 .splice_read = generic_file_splice_read,
563 .splice_write = iter_file_splice_write,
564 .flush = orangefs_flush,
565 .release = orangefs_file_release,
566 .fsync = orangefs_fsync,
567};